[mpich2-commits] r7720 - mpich2/trunk/src/mpi/coll
buntinas at mcs.anl.gov
buntinas at mcs.anl.gov
Thu Jan 13 13:38:44 CST 2011
Author: buntinas
Date: 2011-01-13 13:38:44 -0600 (Thu, 13 Jan 2011)
New Revision: 7720
Modified:
mpich2/trunk/src/mpi/coll/allgather.c
mpich2/trunk/src/mpi/coll/allgatherv.c
mpich2/trunk/src/mpi/coll/allreduce.c
mpich2/trunk/src/mpi/coll/alltoall.c
mpich2/trunk/src/mpi/coll/alltoallw.c
mpich2/trunk/src/mpi/coll/barrier.c
mpich2/trunk/src/mpi/coll/bcast.c
mpich2/trunk/src/mpi/coll/exscan.c
mpich2/trunk/src/mpi/coll/gather.c
mpich2/trunk/src/mpi/coll/gatherv.c
mpich2/trunk/src/mpi/coll/red_scat.c
mpich2/trunk/src/mpi/coll/red_scat_block.c
mpich2/trunk/src/mpi/coll/reduce.c
mpich2/trunk/src/mpi/coll/scan.c
mpich2/trunk/src/mpi/coll/scatter.c
mpich2/trunk/src/mpi/coll/scatterv.c
Log:
Fix collectives to not hang if the communicator contains a failed process. The collectives will not return an error immediately upon detecting a failure, rather they'll return the error at the end of the function and continue the communication pattern so that other processes waiting to receive messages will not hang. This means that, although the collective should complete at all processes, some processes will receive an error, and some processes may not get a valid result. Since some processes may not receive an error and still receive an invalid result, a separate mechanism is needed to confirm that the collective has completed correctly, such as MPI_Comm_validate of the MPI3 FT proposal.
Modified: mpich2/trunk/src/mpi/coll/allgather.c
===================================================================
--- mpich2/trunk/src/mpi/coll/allgather.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/allgather.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -83,11 +83,12 @@
MPI_Datatype recvtype,
MPID_Comm *comm_ptr )
{
- int comm_size, rank;
- int mpi_errno = MPI_SUCCESS;
+ int comm_size, rank;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint recvtype_extent, tot_bytes;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
- int j, i, pof2, src, rem;
+ int j, i, pof2, src, rem;
void *tmp_buf = NULL;
int curr_cnt, dst, type_size, left, right, jnext;
MPI_Comm comm;
@@ -173,11 +174,13 @@
(comm_size-dst_tree_root)*recvcount,
recvtype, dst,
MPIR_ALLGATHER_TAG, comm, &status);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
-
- MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
@@ -235,9 +238,11 @@
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* recv only if this proc. doesn't have data and sender
has data */
@@ -251,10 +256,13 @@
comm, &status);
/* nprocs_completed is also equal to the
no. of processes whose data we don't have */
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
- MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
@@ -331,11 +339,13 @@
tmp_buf_size - recv_offset,
MPI_BYTE, dst,
MPIR_ALLGATHER_TAG, comm, &status);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
-
- MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
@@ -382,8 +392,12 @@
mpi_errno = MPIC_Send(((char *)tmp_buf + offset),
last_recv_cnt, MPI_BYTE,
dst, MPIR_ALLGATHER_TAG,
- comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ comm);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
@@ -398,10 +412,15 @@
MPI_BYTE, dst,
MPIR_ALLGATHER_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
/* nprocs_completed is also equal to the
no. of processes whose data we don't have */
- MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
@@ -469,10 +488,11 @@
curr_cnt, recvtype,
src, MPIR_ALLGATHER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
-
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
curr_cnt *= 2;
pof2 *= 2;
}
@@ -490,9 +510,11 @@
rem * recvcount, recvtype,
src, MPIR_ALLGATHER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Rotate blocks in tmp_buf down by (rank) blocks and store
@@ -549,9 +571,11 @@
recvcount, recvtype, left,
MPIR_ALLGATHER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
j = jnext;
jnext = (comm_size + jnext - 1) % comm_size;
}
@@ -560,8 +584,10 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
- MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
- return (mpi_errno);
+ MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
@@ -590,6 +616,7 @@
*/
int rank, local_size, remote_size, mpi_errno = MPI_SUCCESS, root;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint true_extent, true_lb = 0, extent, send_extent;
void *tmp_buf=NULL;
MPID_Comm *newcomm_ptr = NULL;
@@ -624,9 +651,11 @@
if (sendcount != 0) {
mpi_errno = MPIR_Gather_impl(sendbuf, sendcount, sendtype, tmp_buf, sendcount,
sendtype, 0, newcomm_ptr);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* first broadcast from left to right group, then from right to
@@ -637,9 +666,11 @@
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(tmp_buf, sendcount*local_size,
sendtype, root, comm_ptr);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* receive bcast from right */
@@ -647,9 +678,11 @@
root = 0;
mpi_errno = MPIR_Bcast_inter(recvbuf, recvcount*remote_size,
recvtype, root, comm_ptr);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else {
@@ -658,9 +691,11 @@
root = 0;
mpi_errno = MPIR_Bcast_inter(recvbuf, recvcount*remote_size,
recvtype, root, comm_ptr);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* bcast to left */
@@ -668,14 +703,18 @@
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(tmp_buf, sendcount*local_size,
sendtype, root, comm_ptr);
- if (mpi_errno) {
- MPIU_ERR_POP(mpi_errno);
- }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
- fn_exit:
+ fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
Modified: mpich2/trunk/src/mpi/coll/allgatherv.c
===================================================================
--- mpich2/trunk/src/mpi/coll/allgatherv.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/allgatherv.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -79,7 +79,8 @@
{
MPI_Comm comm;
int comm_size, rank, j, i, left, right;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Aint recvbuf_extent, recvtype_extent, recvtype_true_extent,
recvtype_true_lb;
@@ -191,11 +192,15 @@
total_count - recv_offset, recvtype, dst,
MPIR_ALLGATHERV_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* for convenience, recv is posted for a bigger amount
- than will be sent */
-
- MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ /* for convenience, recv is posted for a bigger amount
+ than will be sent */
+ MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
@@ -254,7 +259,11 @@
last_recv_cnt,
recvtype, dst,
MPIR_ALLGATHERV_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
@@ -273,11 +282,15 @@
total_count - offset, recvtype,
dst, MPIR_ALLGATHERV_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* for convenience, recv is posted for a
- bigger amount than will be sent */
-
- MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ /* for convenience, recv is posted for a
+ bigger amount than will be sent */
+ MPIR_Get_count_impl(&status, recvtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
@@ -377,11 +390,15 @@
((char *)tmp_buf + recv_offset),
tmp_buf_size-recv_offset, MPI_BYTE, dst,
MPIR_ALLGATHERV_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* for convenience, recv is posted for a bigger amount
- than will be sent */
-
- MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ /* for convenience, recv is posted for a bigger amount
+ than will be sent */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
@@ -432,7 +449,11 @@
last_recv_cnt, MPI_BYTE,
dst, MPIR_ALLGATHERV_TAG,
comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* last_recv_cnt was set in the previous
receive. that's the amount of data to be
sent now. */
@@ -447,10 +468,15 @@
dst,
MPIR_ALLGATHERV_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* for convenience, recv is posted for a bigger amount
- than will be sent */
- MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ /* for convenience, recv is posted for a bigger amount
+ than will be sent */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
@@ -523,9 +549,13 @@
((char *)tmp_buf + curr_cnt*recvtype_extent),
total_count - curr_cnt, recvtype,
src, MPIR_ALLGATHERV_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
- MPIR_Get_count_impl(&status, recvtype, &recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, recvtype, &recv_cnt);
curr_cnt += recv_cnt;
pof2 *= 2;
@@ -548,7 +578,11 @@
total_count - curr_cnt, recvtype,
src, MPIR_ALLGATHERV_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Rotate blocks in tmp_buf down by (rank) blocks and store
@@ -631,19 +665,31 @@
}
else if (!sendnow) { /* If there's no data to send, just do a recv call */
mpi_errno = MPIC_Recv(rbuf, recvnow, recvtype, left, MPIR_ALLGATHERV_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
torecv -= recvnow;
}
else if (!recvnow) { /* If there's no data to receive, just do a send call */
mpi_errno = MPIC_Send(sbuf, sendnow, recvtype, right, MPIR_ALLGATHERV_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
tosend -= sendnow;
}
else { /* There's data to be sent and received */
mpi_errno = MPIC_Sendrecv(sbuf, sendnow, recvtype, right, MPIR_ALLGATHERV_TAG,
rbuf, recvnow, recvtype, left, MPIR_ALLGATHERV_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
tosend -= sendnow;
torecv -= recvnow;
}
@@ -665,6 +711,8 @@
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -696,6 +744,7 @@
and then does an intracommunicator broadcast.
*/
int remote_size, mpi_errno, root, rank;
+ int mpi_errno_ret = MPI_SUCCESS;
MPID_Comm *newcomm_ptr = NULL;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
@@ -710,13 +759,21 @@
mpi_errno = MPIR_Gatherv_impl(sendbuf, sendcount, sendtype, recvbuf,
recvcounts, displs, recvtype, root,
comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* gatherv to right group */
root = 0;
mpi_errno = MPIR_Gatherv_impl(sendbuf, sendcount, sendtype, recvbuf,
recvcounts, displs, recvtype, root,
comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* gatherv to left group */
@@ -724,13 +781,21 @@
mpi_errno = MPIR_Gatherv_impl(sendbuf, sendcount, sendtype, recvbuf,
recvcounts, displs, recvtype, root,
comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* gatherv from left group */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Gatherv_impl(sendbuf, sendcount, sendtype, recvbuf,
recvcounts, displs, recvtype, root,
comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* now do an intracommunicator broadcast within each group. we use
@@ -751,11 +816,17 @@
if (mpi_errno) MPIU_ERR_POP(mpi_errno);
mpi_errno = MPIR_Bcast_intra(recvbuf, 1, newtype, 0, newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
MPIR_Type_free_impl(&newtype);
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
/* --BEGIN ERROR HANDLING-- */
Modified: mpich2/trunk/src/mpi/coll/allreduce.c
===================================================================
--- mpich2/trunk/src/mpi/coll/allreduce.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/allreduce.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -131,7 +131,8 @@
int rc;
#endif
int comm_size, rank, type_size;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, pof2, newrank, rem, newdst, i,
send_idx, recv_idx, last_idx, send_cnt, recv_cnt, *cnts, *disps;
MPI_Aint true_extent, true_lb, extent;
@@ -174,10 +175,18 @@
allreduce is in recvbuf. Pass that as the sendbuf to reduce. */
mpi_errno = MPIR_Reduce_impl(recvbuf, NULL, count, datatype, op, 0, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
} else {
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype, op, 0, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
} else {
/* only one process on the node. copy sendbuf to recvbuf */
@@ -190,13 +199,21 @@
/* now do an IN_PLACE allreduce among the local roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = allreduce_intra_or_coll_fn(MPI_IN_PLACE, recvbuf, count, datatype, op, comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* now broadcast the result among local processes */
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
goto fn_exit;
}
@@ -215,17 +232,18 @@
do a reduce to 0 and then broadcast. */
mpi_errno = MPIR_Reduce_impl ( sendbuf, recvbuf, count, datatype,
op, 0, comm_ptr );
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* FIXME: mpi_errno is error CODE, not necessarily the error
- class MPI_ERR_OP. In MPICH2, we can get the error class
- with
- errorclass = mpi_errno & ERROR_CLASS_MASK;
- */
- if (mpi_errno == MPI_ERR_OP || mpi_errno == MPI_SUCCESS) {
- /* Allow MPI_ERR_OP since we can continue from this error */
- rc = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr );
- if (rc) mpi_errno = rc;
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
+
+ mpi_errno = MPIR_Bcast_impl( recvbuf, count, datatype, 0, comm_ptr );
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else
#endif /* MPID_HAS_HETERO */
@@ -299,7 +317,11 @@
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
@@ -311,7 +333,11 @@
datatype, rank-1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
@@ -360,7 +386,11 @@
count, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
@@ -456,7 +486,11 @@
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
@@ -516,7 +550,11 @@
recv_cnt, datatype, dst,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (newrank > newdst) send_idx = recv_idx;
@@ -538,7 +576,11 @@
datatype, rank+1,
MPIR_ALLREDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
if (MPIU_THREADPRIV_FIELD(op_errno))
@@ -550,6 +592,8 @@
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return (mpi_errno);
fn_fail:
@@ -580,6 +624,7 @@
broadcasts because it would require allocation of a temporary buffer.
*/
int rank, mpi_errno, root;
+ int mpi_errno_ret = MPI_SUCCESS;
MPID_Comm *newcomm_ptr = NULL;
rank = comm_ptr->rank;
@@ -591,26 +636,42 @@
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce to rank 0 of right group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* reduce to rank 0 of left group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce from right group to rank 0 */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, recvbuf, count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Get the local intracommunicator */
@@ -620,9 +681,15 @@
newcomm_ptr = comm_ptr->local_comm;
mpi_errno = MPIR_Bcast_impl(recvbuf, count, datatype, 0, newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
Modified: mpich2/trunk/src/mpi/coll/alltoall.c
===================================================================
--- mpich2/trunk/src/mpi/coll/alltoall.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/alltoall.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -88,6 +88,7 @@
MPI_Aint sendtype_extent, recvtype_extent;
MPI_Aint recvtype_true_extent, recvbuf_extent, recvtype_true_lb;
int mpi_errno=MPI_SUCCESS, src, dst, rank, nbytes;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int sendtype_size, pack_size, block, position, *displs, count;
MPI_Datatype newtype = MPI_DATATYPE_NULL;
@@ -138,7 +139,11 @@
j, MPIR_ALLTOALL_TAG,
j, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (rank == j) {
/* same as above with i/j args reversed */
@@ -147,7 +152,11 @@
i, MPIR_ALLTOALL_TAG,
i, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -216,7 +225,11 @@
MPIR_ALLTOALL_TAG, recvbuf, 1, newtype,
src, MPIR_ALLTOALL_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
MPIR_Type_free_impl(&newtype);
@@ -302,11 +315,15 @@
sendbuf_extent*(comm_size-dst_tree_root),
sendtype, dst, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
-
- /* in case of non-power-of-two nodes, less data may be
- received than specified */
- MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ /* in case of non-power-of-two nodes, less data may be
+ received than specified */
+ MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
@@ -351,7 +368,11 @@
last_recv_cnt, sendtype,
dst, MPIR_ALLTOALL_TAG,
comm);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* recv only if this proc. doesn't have data and sender
has data */
@@ -364,8 +385,13 @@
sendtype,
dst, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
- MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ last_recv_cnt = 0;
+ } else
+ MPIR_Get_count_impl(&status, sendtype, &last_recv_cnt);
curr_cnt += last_recv_cnt;
}
tmp_mask >>= 1;
@@ -430,7 +456,7 @@
recvcount, recvtype, dst,
MPIR_ALLTOALL_TAG, comm,
&reqarray[i]);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
for ( i=0; i<ss; i++ ) {
@@ -440,7 +466,7 @@
sendcount, sendtype, dst,
MPIR_ALLTOALL_TAG, comm,
&reqarray[i+ss]);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) MPIU_ERR_POP(mpi_errno);
}
/* ... then wait for them to finish: */
@@ -452,7 +478,11 @@
for (j=0; j<2*ss; j++) {
if (starray[j].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[j].MPI_ERROR;
- MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -502,7 +532,11 @@
src*recvcount*recvtype_extent),
recvcount, recvtype, src,
MPIR_ALLTOALL_TAG, comm, &status);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -510,7 +544,9 @@
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
if (newtype != MPI_DATATYPE_NULL)
MPIR_Type_free_impl(&newtype);
@@ -544,7 +580,8 @@
*/
int local_size, remote_size, max_size, i;
MPI_Aint sendtype_extent, recvtype_extent;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int src, dst, rank;
char *sendaddr, *recvaddr;
@@ -590,13 +627,19 @@
MPIR_ALLTOALL_TAG, recvaddr,
recvcount, recvtype, src,
MPIR_ALLTOALL_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
Modified: mpich2/trunk/src/mpi/coll/alltoallw.c
===================================================================
--- mpich2/trunk/src/mpi/coll/alltoallw.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/alltoallw.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -62,7 +62,8 @@
MPID_Comm *comm_ptr )
{
int comm_size, i, j;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Status *starray;
MPI_Request *reqarray;
@@ -100,7 +101,11 @@
j, MPIR_ALLTOALL_TAG,
j, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (rank == j) {
/* same as above with i/j args reversed */
@@ -109,7 +114,11 @@
i, MPIR_ALLTOALL_TAG,
i, MPIR_ALLTOALL_TAG,
comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -167,7 +176,11 @@
for (i=0; i<outstanding_requests; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
- MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -193,7 +206,11 @@
((char *)recvbuf+rdispls[src]),
recvcnts[src], recvtypes[dst], src,
MPIR_ALLTOALLW_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
#endif
}
@@ -202,7 +219,9 @@
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
@@ -238,6 +257,7 @@
*/
int local_size, remote_size, max_size, i;
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int src, dst, rank, sendcount, recvcount;
char *sendaddr, *recvaddr;
@@ -284,13 +304,19 @@
dst, MPIR_ALLTOALLW_TAG, recvaddr,
recvcount, recvtype, src,
MPIR_ALLTOALLW_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
fn_exit:
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
Modified: mpich2/trunk/src/mpi/coll/barrier.c
===================================================================
--- mpich2/trunk/src/mpi/coll/barrier.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/barrier.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -55,6 +55,7 @@
int MPIR_Barrier_intra( MPID_Comm *comm_ptr )
{
int size, rank, src, dst, mask, mpi_errno=MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
/* Only one collective operation per communicator can be active at any
@@ -76,13 +77,18 @@
MPIR_BARRIER_TAG, NULL, 0, MPI_BYTE,
src, MPIR_BARRIER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
mask <<= 1;
}
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -128,6 +134,7 @@
int MPIR_Barrier_inter( MPID_Comm *comm_ptr )
{
int rank, mpi_errno = MPI_SUCCESS, root;
+ int mpi_errno_ret = MPI_SUCCESS;
int i = 0;
MPID_Comm *newcomm_ptr = NULL;
@@ -143,7 +150,11 @@
/* do a barrier on the local intracommunicator */
mpi_errno = MPIR_Barrier_intra(newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* rank 0 on each group does an intercommunicator broadcast to the
remote group to indicate that all processes in the local group
@@ -156,23 +167,41 @@
/* bcast to right*/
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* receive bcast from right */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* receive bcast from left */
root = 0;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* bcast to left */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Bcast_inter(&i, 1, MPI_BYTE, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -217,6 +246,7 @@
int MPIR_Barrier_impl(MPID_Comm *comm_ptr)
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
if (comm_ptr->coll_fns != NULL && comm_ptr->coll_fns->Barrier != NULL)
{
@@ -233,13 +263,21 @@
if (comm_ptr->node_comm != NULL)
{
mpi_errno = MPIR_Barrier_or_coll_fn(comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* do the barrier across roots of all nodes */
if (comm_ptr->node_roots_comm != NULL) {
mpi_errno = MPIR_Barrier_or_coll_fn(comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* release the local processes on each node with a 1-byte broadcast
@@ -248,7 +286,11 @@
{
int i=0;
mpi_errno = MPIR_Bcast_impl(&i, 1, MPI_BYTE, 0, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else {
@@ -268,6 +310,8 @@
}
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
Modified: mpich2/trunk/src/mpi/coll/bcast.c
===================================================================
--- mpich2/trunk/src/mpi/coll/bcast.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/bcast.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -58,7 +58,8 @@
{
int rank, comm_size, src, dst;
int relative_rank, mask;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int nbytes=0;
int type_size, is_contig, is_homogeneous;
int position;
@@ -154,7 +155,11 @@
else
mpi_errno = MPIC_Recv(buffer,count,datatype,src,
MPIR_BCAST_TAG,comm,MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
break;
}
mask <<= 1;
@@ -184,7 +189,11 @@
else
mpi_errno = MPIC_Send(buffer,count,datatype,dst,
MPIR_BCAST_TAG,comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
mask >>= 1;
}
@@ -203,6 +212,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -236,7 +247,8 @@
MPI_Status status;
int rank, comm_size, src, dst;
int relative_rank, mask;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int scatter_size, curr_size, recv_size = 0, send_size;
MPI_Comm comm;
@@ -283,10 +295,14 @@
relative_rank*scatter_size),
recv_size, MPI_BYTE, src,
MPIR_BCAST_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
- /* query actual size of data received */
- MPIR_Get_count_impl(&status, MPI_BYTE, &curr_size);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ curr_size = 0;
+ } else
+ /* query actual size of data received */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &curr_size);
}
break;
}
@@ -314,7 +330,11 @@
scatter_size*(relative_rank+mask)),
send_size, MPI_BYTE, dst,
MPIR_BCAST_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
curr_size -= send_size;
}
@@ -323,6 +343,8 @@
}
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -363,6 +385,7 @@
int rank, comm_size, dst;
int relative_rank, mask;
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int scatter_size, nbytes=0, curr_size, recv_size = 0;
int type_size, j, k, i, tmp_mask, is_contig, is_homogeneous;
int relative_dst, dst_tree_root, my_tree_root, send_offset;
@@ -436,7 +459,11 @@
mpi_errno = scatter_for_bcast(buffer, count, datatype, root, comm_ptr,
nbytes, tmp_buf, is_contig, is_homogeneous);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* medium size allgather and pof2 comm_size. use recurive doubling. */
@@ -470,9 +497,13 @@
((char *)tmp_buf + recv_offset),
(nbytes-recv_offset < 0 ? 0 : nbytes-recv_offset),
MPI_BYTE, dst, MPIR_BCAST_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
- MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ recv_size = 0;
+ } else
+ MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
curr_size += recv_size;
}
@@ -540,7 +571,11 @@
/* recv_size was set in the previous
receive. that's the amount of data to be
sent now. */
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* recv only if this proc. doesn't have data and sender
has data */
@@ -556,9 +591,13 @@
comm, &status);
/* nprocs_completed is also equal to the no. of processes
whose data we don't have */
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
- MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ recv_size = 0;
+ } else
+ MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
curr_size += recv_size;
/* printf("Rank %d, recv from %d, offset %d, size %d\n", rank, dst, offset, recv_size);
fflush(stdout);*/
@@ -586,6 +625,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -621,6 +662,7 @@
int rank, comm_size;
int relative_rank;
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int scatter_size, nbytes;
int type_size, j, i, is_contig, is_homogeneous;
int position;
@@ -690,7 +732,11 @@
mpi_errno = scatter_for_bcast(buffer, count, datatype, root, comm_ptr,
nbytes, tmp_buf, is_contig, is_homogeneous);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* long-message allgather or medium-size but non-power-of-two. use ring algorithm. */
@@ -727,7 +773,11 @@
recvcnts[(jnext-root+comm_size)%comm_size],
MPI_BYTE, left,
MPIR_BCAST_TAG, comm, MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
j = jnext;
jnext = (comm_size + jnext - 1) % comm_size;
@@ -746,6 +796,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -769,7 +821,11 @@
{ \
mpi_errno_ = bcast_fn_(buffer_, count_, datatype_, root_, comm_ptr_); \
} \
- if (mpi_errno_) MPIU_ERR_POP(mpi_errno_); \
+ if (mpi_errno) { \
+ /* for communication errors, just record the error but continue */ \
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail"); \
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno); \
+ } \
} while (0)
/* FIXME This function uses some heuristsics based off of some testing on a
@@ -786,6 +842,7 @@
MPID_Comm *comm_ptr)
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int type_size, is_homogeneous;
int nbytes=0;
@@ -828,7 +885,11 @@
mpi_errno = MPIC_Recv(buffer,count,datatype,MPIU_Get_intranode_rank(comm_ptr, root),
MPIR_BCAST_TAG,comm_ptr->node_comm->handle,MPI_STATUS_IGNORE);
}
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* perform the internode broadcast */
@@ -905,11 +966,17 @@
algorithm that (at least approximately) minimized internode
communication. */
mpi_errno = MPIR_Bcast_scatter_ring_allgather(buffer, count, datatype, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
fn_exit:
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -971,6 +1038,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int comm_size;
int nbytes=0;
int type_size, is_homogeneous;
@@ -987,7 +1055,11 @@
#if defined(USE_SMP_COLLECTIVES)
if (MPIR_Comm_is_node_aware(comm_ptr)) {
mpi_errno = MPIR_SMP_Bcast(buffer, count, datatype, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
goto fn_exit;
}
#endif
@@ -1018,14 +1090,22 @@
if ((nbytes < MPIR_PARAM_BCAST_SHORT_MSG_SIZE) || (comm_size < MPIR_PARAM_BCAST_MIN_PROCS))
{
mpi_errno = MPIR_Bcast_binomial(buffer, count, datatype, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else /* (nbytes >= MPIR_PARAM_BCAST_SHORT_MSG_SIZE) && (comm_size >= MPIR_PARAM_BCAST_MIN_PROCS) */
{
if ((nbytes < MPIR_PARAM_BCAST_LONG_MSG_SIZE) && (MPIU_is_pof2(comm_size, NULL)))
{
mpi_errno = MPIR_Bcast_scatter_doubling_allgather(buffer, count, datatype, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else /* (nbytes >= MPIR_PARAM_BCAST_LONG_MSG_SIZE) || !(comm_size_is_pof2) */
{
@@ -1033,7 +1113,11 @@
topologically aware communicator. Doing inter/intra-node
communication phases breaks the pipelining of the algorithm. */
mpi_errno = MPIR_Bcast_scatter_ring_allgather(buffer, count, datatype, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -1043,6 +1127,8 @@
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_BCAST);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -1067,6 +1153,7 @@
intracommunicator broadcast.
*/
int rank, mpi_errno;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPID_Comm *newcomm_ptr = NULL;
MPI_Comm comm;
@@ -1087,7 +1174,11 @@
MPIDU_ERR_CHECK_MULTIPLE_THREADS_ENTER( comm_ptr );
mpi_errno = MPIC_Send(buffer, count, datatype, 0,
MPIR_BCAST_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
}
else
@@ -1100,7 +1191,11 @@
{
mpi_errno = MPIC_Recv(buffer, count, datatype, root,
MPIR_BCAST_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Get the local intracommunicator */
@@ -1112,11 +1207,17 @@
/* now do the usual broadcast on this intracommunicator
with rank 0 as root. */
mpi_errno = MPIR_Bcast_intra(buffer, count, datatype, 0, newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
fn_fail:
MPID_MPI_FUNC_EXIT(MPID_STATE_MPIR_BCAST_INTER);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
}
Modified: mpich2/trunk/src/mpi/coll/exscan.c
===================================================================
--- mpich2/trunk/src/mpi/coll/exscan.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/exscan.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -106,7 +106,8 @@
{
MPI_Status status;
int rank, comm_size;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative, flag;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
@@ -187,7 +188,11 @@
count, datatype, dst,
MPIR_EXSCAN_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (rank > dst) {
call_uop(tmp_buf, partial_scan, count, datatype);
@@ -237,7 +242,9 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
Modified: mpich2/trunk/src/mpi/coll/gather.c
===================================================================
--- mpich2/trunk/src/mpi/coll/gather.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/gather.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -61,7 +61,8 @@
MPID_Comm *comm_ptr )
{
int comm_size, rank;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int curr_cnt=0, relative_rank, nbytes, is_homogeneous;
int mask, sendtype_size, recvtype_size, src, dst, relative_src;
int recvblks;
@@ -194,7 +195,11 @@
recvblks * recvcnt, recvtype, src,
MPIR_GATHER_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (nbytes < MPIR_PARAM_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIC_Recv(tmp_buf, recvblks * nbytes, MPI_BYTE,
@@ -218,8 +223,9 @@
mpi_errno = MPIC_Recv(recvbuf, 1, tmp_type, src,
MPIR_GATHER_TAG, comm, &status);
if (mpi_errno) {
- MPIR_Type_free_impl(&tmp_type);
- MPIU_ERR_POP(mpi_errno);
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIR_Type_free_impl(&tmp_type);
@@ -243,7 +249,11 @@
recvblks * nbytes, MPI_BYTE, src,
MPIR_GATHER_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
curr_cnt += (recvblks * nbytes);
}
}
@@ -258,12 +268,20 @@
/* leaf nodes send directly from sendbuf */
mpi_errno = MPIC_Send(sendbuf, sendcnt, sendtype, dst,
MPIR_GATHER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (nbytes < MPIR_PARAM_GATHER_VSMALL_MSG_SIZE) {
mpi_errno = MPIC_Send(tmp_buf, curr_cnt, MPI_BYTE, dst,
MPIR_GATHER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
blocks[0] = sendcnt;
@@ -282,8 +300,9 @@
mpi_errno = MPIC_Send(MPI_BOTTOM, 1, tmp_type, dst,
MPIR_GATHER_TAG, comm);
if (mpi_errno) {
- MPIR_Type_free_impl(&tmp_type);
- MPIU_ERR_POP(mpi_errno);
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
}
MPIR_Type_free_impl(&tmp_type);
}
@@ -352,10 +371,15 @@
tmp_buf_size-curr_cnt, MPI_BYTE, src,
MPIR_GATHER_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* the recv size is larger than what may be sent in
- some cases. query amount of data actually received */
- MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ recv_size = 0;
+ } else
+ /* the recv size is larger than what may be sent in
+ some cases. query amount of data actually received */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &recv_size);
curr_cnt += recv_size;
}
}
@@ -365,7 +389,11 @@
dst = (dst + root) % comm_size;
mpi_errno = MPIC_Send(tmp_buf, curr_cnt, MPI_BYTE, dst,
MPIR_GATHER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
break;
}
mask <<= 1;
@@ -404,6 +432,8 @@
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -438,6 +468,7 @@
*/
int rank, local_size, remote_size, mpi_errno=MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int i, nbytes, sendtype_size, recvtype_size;
MPI_Status status;
MPI_Aint extent, true_extent, true_lb = 0;
@@ -478,8 +509,11 @@
mpi_errno = MPIC_Recv(recvbuf, recvcnt*remote_size,
recvtype, 0, MPIR_GATHER_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else
{
@@ -513,14 +547,22 @@
mpi_errno = MPIR_Gather_impl(sendbuf, sendcnt, sendtype,
tmp_buf, sendcnt, sendtype, 0,
newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (rank == 0)
{
mpi_errno = MPIC_Send(tmp_buf, sendcnt*local_size,
sendtype, root,
MPIR_GATHER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -538,20 +580,30 @@
mpi_errno = MPIC_Recv(((char *)recvbuf+recvcnt*i*extent),
recvcnt, recvtype, i,
MPIR_GATHER_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else
{
mpi_errno = MPIC_Send(sendbuf,sendcnt,sendtype,root,
MPIR_GATHER_TAG,comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
fn_exit:
MPIU_CHKLMEM_FREEALL();
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
Modified: mpich2/trunk/src/mpi/coll/gatherv.c
===================================================================
--- mpich2/trunk/src/mpi/coll/gatherv.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/gatherv.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -61,6 +61,7 @@
{
int comm_size, rank;
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
MPI_Aint extent;
int i, reqs;
@@ -120,7 +121,11 @@
for (i = 0; i < reqs; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
- MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -143,12 +148,20 @@
if (comm_size >= min_procs) {
mpi_errno = MPIC_Ssend(sendbuf, sendcnt, sendtype, root,
MPIR_GATHERV_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
mpi_errno = MPIC_Send(sendbuf, sendcnt, sendtype, root,
MPIR_GATHERV_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -158,6 +171,8 @@
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
Modified: mpich2/trunk/src/mpi/coll/red_scat.c
===================================================================
--- mpich2/trunk/src/mpi/coll/red_scat.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/red_scat.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -83,6 +83,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int comm_size = comm_ptr->local_size;
int rank = comm_ptr->rank;
int pof2;
@@ -191,7 +192,11 @@
incoming_data + recv_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_TAG,
comm, MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* always perform the reduction at recv_offset, the data at send_offset
is now our peer's responsibility */
if (rank > peer) {
@@ -222,6 +227,8 @@
recvbuf, size, datatype);
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -292,7 +299,8 @@
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int type_size, dis[2], blklens[2], total_count, nbytes, src, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
@@ -408,7 +416,11 @@
mpi_errno = MPIC_Send(tmp_results, total_count,
datatype, rank+1,
MPIR_REDUCE_SCATTER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
@@ -420,7 +432,11 @@
datatype, rank-1,
MPIR_REDUCE_SCATTER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
@@ -519,7 +535,11 @@
dst, MPIR_REDUCE_SCATTER_TAG,
comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
@@ -567,7 +587,11 @@
disps[rank-1]*extent, recvcnts[rank-1],
datatype, rank-1,
MPIR_REDUCE_SCATTER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else { /* even */
@@ -576,7 +600,11 @@
datatype, rank+1,
MPIR_REDUCE_SCATTER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -621,7 +649,11 @@
MPIR_REDUCE_SCATTER_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (is_commutative || (src < rank)) {
if (sendbuf != MPI_IN_PLACE) {
@@ -818,7 +850,11 @@
MPIR_REDUCE_SCATTER_TAG, comm,
MPI_STATUS_IGNORE);
received = 1;
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* if some processes in this process's subtree in this step
@@ -871,7 +907,11 @@
MPIR_REDUCE_SCATTER_TAG,
comm, MPI_STATUS_IGNORE);
received = 1;
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
tmp_mask >>= 1;
k--;
@@ -959,7 +999,9 @@
if (MPIU_THREADPRIV_FIELD(op_errno))
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
@@ -986,6 +1028,7 @@
*/
int rank, mpi_errno, root, local_size, total_count, i;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint true_extent, true_lb = 0, extent;
void *tmp_buf=NULL;
int *disps=NULL;
@@ -1026,26 +1069,42 @@
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce to rank 0 of right group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* reduce to rank 0 of left group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce from right group to rank 0 */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Get the local intracommunicator */
@@ -1058,10 +1117,16 @@
mpi_errno = MPIR_Scatterv(tmp_buf, recvcnts, disps, datatype, recvbuf,
recvcnts[rank], datatype, 0, newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -1073,7 +1138,7 @@
implementations of reduce_scatter. In all other cases
MPIR_Reduce_Scatter_impl should be used. */
#undef FUNCNAME
-#define FUNCNAME MPIR_Reduce_scatter_impl
+#define FUNCNAME MPIR_Reduce_scatter
#undef FCNAME
#define FCNAME MPIU_QUOTE(FUNCNAME)
int MPIR_Reduce_scatter(void *sendbuf, void *recvbuf, int *recvcnts,
Modified: mpich2/trunk/src/mpi/coll/red_scat_block.c
===================================================================
--- mpich2/trunk/src/mpi/coll/red_scat_block.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/red_scat_block.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -89,6 +89,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int comm_size = comm_ptr->local_size;
int rank = comm_ptr->rank;
int pof2;
@@ -193,7 +194,11 @@
incoming_data + recv_offset*true_extent,
size, datatype, peer, MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm, MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* always perform the reduction at recv_offset, the data at send_offset
is now our peer's responsibility */
if (rank > peer) {
@@ -226,6 +231,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -296,7 +303,8 @@
MPI_Aint extent, true_extent, true_lb;
int *disps;
void *tmp_recvbuf, *tmp_results;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int type_size, dis[2], blklens[2], total_count, nbytes, src, dst;
int mask, dst_tree_root, my_tree_root, j, k;
int *newcnts, *newdisps, rem, newdst, send_idx, recv_idx,
@@ -411,7 +419,11 @@
mpi_errno = MPIC_Send(tmp_results, total_count,
datatype, rank+1,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
@@ -423,7 +435,11 @@
datatype, rank-1,
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* do the reduction on received data. since the
ordering is right, it doesn't matter whether
@@ -522,7 +538,11 @@
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* tmp_recvbuf contains data received in this step.
tmp_results contains data accumulated so far */
@@ -573,7 +593,11 @@
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE);
}
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -616,7 +640,11 @@
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (is_commutative || (src < rank)) {
if (sendbuf != MPI_IN_PLACE) {
@@ -803,7 +831,11 @@
MPIR_REDUCE_SCATTER_BLOCK_TAG, comm,
MPI_STATUS_IGNORE);
received = 1;
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* if some processes in this process's subtree in this step
@@ -845,7 +877,11 @@
mpi_errno = MPIC_Send(tmp_recvbuf, 1, recvtype,
dst, MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* recv only if this proc. doesn't have data and sender
has data */
@@ -856,7 +892,11 @@
MPIR_REDUCE_SCATTER_BLOCK_TAG,
comm, MPI_STATUS_IGNORE);
received = 1;
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
tmp_mask >>= 1;
k--;
@@ -944,7 +984,9 @@
if (MPIU_THREADPRIV_FIELD(op_errno))
mpi_errno = MPIU_THREADPRIV_FIELD(op_errno);
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
@@ -971,6 +1013,7 @@
*/
int rank, mpi_errno, root, local_size, total_count;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Aint true_extent, true_lb = 0, extent;
void *tmp_buf=NULL;
MPID_Comm *newcomm_ptr = NULL;
@@ -1001,26 +1044,42 @@
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce to rank 0 of right group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* reduce to rank 0 of left group */
root = 0;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* reduce from right group to rank 0 */
root = (rank == 0) ? MPI_ROOT : MPI_PROC_NULL;
mpi_errno = MPIR_Reduce_inter(sendbuf, tmp_buf, total_count, datatype, op,
root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Get the local intracommunicator */
@@ -1031,10 +1090,16 @@
mpi_errno = MPIR_Scatter_impl(tmp_buf, recvcount, datatype, recvbuf,
recvcount, datatype, 0, newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
Modified: mpich2/trunk/src/mpi/coll/reduce.c
===================================================================
--- mpich2/trunk/src/mpi/coll/reduce.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/reduce.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -41,6 +41,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
int comm_size, rank, is_commutative, type_size;
int mask, relrank, source, lroot;
@@ -168,7 +169,11 @@
source = (source + lroot) % comm_size;
mpi_errno = MPIC_Recv (tmp_buf, count, datatype, source,
MPIR_REDUCE_TAG, comm, &status);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* The sender is above us, so the received buffer must be
the second argument (in the noncommutative case). */
@@ -203,7 +208,11 @@
source = ((relrank & (~ mask)) + lroot) % comm_size;
mpi_errno = MPIC_Send( recvbuf, count, datatype,
source, MPIR_REDUCE_TAG, comm );
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
break;
}
mask <<= 1;
@@ -221,7 +230,11 @@
mpi_errno = MPIC_Recv ( recvbuf, count, datatype, 0,
MPIR_REDUCE_TAG, comm, &status);
}
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* FIXME does this need to be checked after each uop invocation for
@@ -235,6 +248,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -281,6 +296,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int comm_size, rank, is_commutative, type_size, pof2, rem, newrank;
int mask, *cnts, *disps, i, j, send_idx=0;
int recv_idx, last_idx=0, newdst;
@@ -389,7 +405,11 @@
mpi_errno = MPIC_Send(recvbuf, count,
datatype, rank-1,
MPIR_REDUCE_TAG, comm);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* temporarily set the rank to -1 so that this
process does not pariticipate in recursive
@@ -401,7 +421,11 @@
datatype, rank+1,
MPIR_REDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* do the reduction on received data. */
/* This algorithm is used only for predefined ops
@@ -480,7 +504,11 @@
recv_cnt, datatype, dst,
MPIR_REDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* tmp_buf contains data received in this step.
recvbuf contains data accumulated so far */
@@ -534,7 +562,11 @@
mpi_errno = MPIC_Recv(recvbuf, cnts[0], datatype,
0, MPIR_REDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
newrank = 0;
send_idx = 0;
last_idx = 2;
@@ -542,7 +574,11 @@
else if (newrank == 0) { /* send */
mpi_errno = MPIC_Send(recvbuf, cnts[0], datatype,
root, MPIR_REDUCE_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
newrank = -1;
}
newroot = 0;
@@ -611,7 +647,11 @@
send_cnt, datatype,
dst, MPIR_REDUCE_TAG,
comm);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
break;
}
else {
@@ -623,7 +663,11 @@
recv_cnt, datatype, dst,
MPIR_REDUCE_TAG, comm,
MPI_STATUS_IGNORE);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
if (newrank > newdst) send_idx = recv_idx;
@@ -644,6 +688,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -720,6 +766,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int comm_size, is_commutative, type_size, pof2;
MPID_Op *op_ptr;
#if defined(USE_SMP_COLLECTIVES)
@@ -762,7 +809,11 @@
MPIU_Get_intranode_rank(comm_ptr, root) == -1) {
mpi_errno = MPIR_Reduce_impl(sendbuf, tmp_buf, count, datatype,
op, 0, comm_ptr->node_comm);
- if (mpi_errno) goto fn_fail;
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* do the internode reduce to the root's node */
@@ -774,7 +825,11 @@
mpi_errno = MPIR_Reduce_impl(buf, NULL, count, datatype,
op, MPIU_Get_internode_rank(comm_ptr, root),
comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else { /* I am on root's node. I have not participated in the earlier reduce. */
if (comm_ptr->rank != root) {
@@ -784,7 +839,11 @@
mpi_errno = MPIR_Reduce_impl(sendbuf, tmp_buf, count, datatype,
op, MPIU_Get_internode_rank(comm_ptr, root),
comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* point sendbuf at tmp_buf to make final intranode reduce easy */
sendbuf = tmp_buf;
@@ -795,7 +854,11 @@
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype,
op, MPIU_Get_internode_rank(comm_ptr, root),
comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
/* set sendbuf to MPI_IN_PLACE to make final intranode reduce easy. */
sendbuf = MPI_IN_PLACE;
@@ -810,7 +873,11 @@
mpi_errno = MPIR_Reduce_impl(sendbuf, recvbuf, count, datatype,
op, MPIU_Get_intranode_rank(comm_ptr, root),
comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
goto fn_exit;
@@ -842,12 +909,20 @@
(HANDLE_GET_KIND(op) == HANDLE_KIND_BUILTIN) && (count >= pof2)) {
/* do a reduce-scatter followed by gather to root. */
mpi_errno = MPIR_Reduce_redscat_gather(sendbuf, recvbuf, count, datatype, op, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* use a binomial tree algorithm */
mpi_errno = MPIR_Reduce_binomial(sendbuf, recvbuf, count, datatype, op, root, comm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
@@ -857,6 +932,8 @@
#if defined(USE_SMP_COLLECTIVES)
MPIU_CHKLMEM_FREEALL();
#endif
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -886,6 +963,7 @@
*/
int rank, mpi_errno;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Status status;
MPI_Aint true_extent, true_lb, extent;
void *tmp_buf=NULL;
@@ -906,7 +984,11 @@
/* root receives data from rank 0 on remote group */
mpi_errno = MPIC_Recv(recvbuf, count, datatype, 0,
MPIR_REDUCE_TAG, comm, &status);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
/* remote group. Rank 0 allocates temporary buffer, does
@@ -939,19 +1021,29 @@
/* now do a local reduce on this intracommunicator */
mpi_errno = MPIR_Reduce_intra(sendbuf, tmp_buf, count, datatype,
op, 0, newcomm_ptr);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (rank == 0)
{
mpi_errno = MPIC_Send(tmp_buf, count, datatype, root,
MPIR_REDUCE_TAG, comm);
- if (mpi_errno) { MPIU_ERR_POP(mpi_errno); }
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
fn_exit:
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
Modified: mpich2/trunk/src/mpi/coll/scan.c
===================================================================
--- mpich2/trunk/src/mpi/coll/scan.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/scan.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -75,7 +75,8 @@
{
MPI_Status status;
int rank, comm_size;
- int mpi_errno = MPI_SUCCESS;
+ int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int mask, dst, is_commutative;
MPI_Aint true_extent, true_lb, extent;
void *partial_scan, *tmp_buf;
@@ -171,7 +172,11 @@
count, datatype, dst,
MPIR_SCAN_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (rank > dst) {
#ifdef HAVE_CXX_BINDING
@@ -228,7 +233,9 @@
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
- return (mpi_errno);
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
+ return mpi_errno;
fn_fail:
goto fn_exit;
}
@@ -252,6 +259,7 @@
MPID_Comm *comm_ptr )
{
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPIU_CHKLMEM_DECL(3);
MPIU_THREADPRIV_DECL;
int rank = comm_ptr->rank;
@@ -303,7 +311,11 @@
{
mpi_errno = MPIR_Scan_impl(sendbuf, recvbuf, count, datatype,
op, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (sendbuf != MPI_IN_PLACE)
{
@@ -321,7 +333,11 @@
mpi_errno = MPIC_Recv(localfulldata, count, datatype,
comm_ptr->node_comm->local_size - 1, MPIR_SCAN_TAG,
comm_ptr->node_comm->handle, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (comm_ptr->node_roots_comm == NULL &&
comm_ptr->node_comm != NULL &&
@@ -329,7 +345,11 @@
{
mpi_errno = MPIC_Send(recvbuf, count, datatype,
0, MPIR_SCAN_TAG, comm_ptr->node_comm->handle);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else if (comm_ptr->node_roots_comm != NULL)
{
@@ -344,7 +364,11 @@
{
mpi_errno = MPIR_Scan_impl(localfulldata, prefulldata, count, datatype,
op, comm_ptr->node_roots_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
if (MPIU_Get_internode_rank(comm_ptr, rank) !=
comm_ptr->node_roots_comm->local_size-1)
@@ -352,7 +376,11 @@
mpi_errno = MPIC_Send(prefulldata, count, datatype,
MPIU_Get_internode_rank(comm_ptr, rank) + 1,
MPIR_SCAN_TAG, comm_ptr->node_roots_comm->handle);
- if(mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
if (MPIU_Get_internode_rank(comm_ptr, rank) != 0)
{
@@ -361,7 +389,11 @@
MPIR_SCAN_TAG, comm_ptr->node_roots_comm->handle,
&status);
noneed = 0;
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -373,7 +405,11 @@
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(&noneed, 1, MPI_INT, 0, comm_ptr->node_comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
if (noneed == 0) {
@@ -382,7 +418,11 @@
#endif
if (comm_ptr->node_comm != NULL) {
mpi_errno = MPIR_Bcast_impl(tempbuf, count, datatype, 0, comm_ptr->node_comm);
- if(mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* do reduce on tempbuf and recvbuf, finish scan. */
@@ -420,6 +460,8 @@
fn_exit:
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
Modified: mpich2/trunk/src/mpi/coll/scatter.c
===================================================================
--- mpich2/trunk/src/mpi/coll/scatter.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/scatter.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -68,6 +68,7 @@
int tmp_buf_size = 0;
void *tmp_buf=NULL;
int mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
MPIU_CHKLMEM_DECL(4);
@@ -171,16 +172,24 @@
mpi_errno = MPIC_Recv(recvbuf, recvcnt, recvtype,
src, MPIR_SCATTER_TAG, comm,
&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
else {
mpi_errno = MPIC_Recv(tmp_buf, tmp_buf_size, MPI_BYTE, src,
MPIR_SCATTER_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
-
- /* the recv size is larger than what may be sent in
- some cases. query amount of data actually received */
- MPIR_Get_count_impl(&status, MPI_BYTE, &curr_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ curr_cnt = 0;
+ } else
+ /* the recv size is larger than what may be sent in
+ some cases. query amount of data actually received */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &curr_cnt);
}
break;
}
@@ -218,7 +227,11 @@
MPI_BYTE, dst,
MPIR_SCATTER_TAG, comm);
}
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
curr_cnt -= send_subtree_cnt;
}
mask >>= 1;
@@ -319,10 +332,15 @@
mpi_errno = MPIC_Recv(tmp_buf, tmp_buf_size, MPI_BYTE, src,
MPIR_SCATTER_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
- /* the recv size is larger than what may be sent in
- some cases. query amount of data actually received */
- MPIR_Get_count_impl(&status, MPI_BYTE, &curr_cnt);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ curr_cnt = 0;
+ } else
+ /* the recv size is larger than what may be sent in
+ some cases. query amount of data actually received */
+ MPIR_Get_count_impl(&status, MPI_BYTE, &curr_cnt);
break;
}
mask <<= 1;
@@ -344,7 +362,11 @@
mpi_errno = MPIC_Send (((char *)tmp_buf + nbytes*mask),
send_subtree_cnt, MPI_BYTE, dst,
MPIR_SCATTER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
curr_cnt -= send_subtree_cnt;
}
mask >>= 1;
@@ -364,6 +386,8 @@
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
@@ -396,6 +420,7 @@
*/
int rank, local_size, remote_size, mpi_errno=MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
int i, nbytes, sendtype_size, recvtype_size;
MPI_Status status;
MPI_Aint extent, true_extent, true_lb = 0;
@@ -429,7 +454,11 @@
/* root sends all data to rank 0 on remote group and returns */
mpi_errno = MPIC_Send(sendbuf, sendcnt*remote_size,
sendtype, 0, MPIR_SCATTER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
goto fn_exit;
}
else {
@@ -454,7 +483,11 @@
mpi_errno = MPIC_Recv(tmp_buf, recvcnt*local_size,
recvtype, root,
MPIR_SCATTER_TAG, comm, &status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
/* Get the local intracommunicator */
@@ -467,7 +500,11 @@
mpi_errno = MPIR_Scatter_impl(tmp_buf, recvcnt, recvtype,
recvbuf, recvcnt, recvtype, 0,
newcomm_ptr);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else {
@@ -478,13 +515,21 @@
mpi_errno = MPIC_Send(((char *)sendbuf+sendcnt*i*extent),
sendcnt, sendtype, i,
MPIR_SCATTER_TAG, comm);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
else {
mpi_errno = MPIC_Recv(recvbuf,recvcnt,recvtype,root,
MPIR_SCATTER_TAG,comm,&status);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -492,6 +537,8 @@
MPIU_CHKLMEM_FREEALL();
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
Modified: mpich2/trunk/src/mpi/coll/scatterv.c
===================================================================
--- mpich2/trunk/src/mpi/coll/scatterv.c 2011-01-13 19:28:01 UTC (rev 7719)
+++ mpich2/trunk/src/mpi/coll/scatterv.c 2011-01-13 19:38:44 UTC (rev 7720)
@@ -60,6 +60,7 @@
MPID_Comm *comm_ptr )
{
int rank, comm_size, mpi_errno = MPI_SUCCESS;
+ int mpi_errno_ret = MPI_SUCCESS;
MPI_Comm comm;
MPI_Aint extent;
int i, reqs;
@@ -120,7 +121,11 @@
for (i = 0; i < reqs; i++) {
if (starray[i].MPI_ERROR != MPI_SUCCESS) {
mpi_errno = starray[i].MPI_ERROR;
- MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
}
@@ -131,7 +136,11 @@
if (recvcnt) {
mpi_errno = MPIC_Recv(recvbuf,recvcnt,recvtype,root,
MPIR_SCATTERV_TAG,comm,MPI_STATUS_IGNORE);
- if (mpi_errno) MPIU_ERR_POP(mpi_errno);
+ if (mpi_errno) {
+ /* for communication errors, just record the error but continue */
+ MPIU_ERR_SET(mpi_errno, MPI_ERR_OTHER, "**fail");
+ MPIU_ERR_ADD(mpi_errno_ret, mpi_errno);
+ }
}
}
@@ -140,6 +149,8 @@
/* check if multiple threads are calling this collective function */
MPIDU_ERR_CHECK_MULTIPLE_THREADS_EXIT( comm_ptr );
MPIU_CHKLMEM_FREEALL();
+ if (mpi_errno_ret)
+ mpi_errno = mpi_errno_ret;
return mpi_errno;
fn_fail:
goto fn_exit;
More information about the mpich2-commits
mailing list